Nah, I was experimenting with an entirely separate lattice. Regular inference was already taking care of Const and ParticalConst. The problem was that the abstract interpreter, at the time, did NOT like having to lug around that separate lattice while also doing its own inference, and the docs about how to extend that were (are?) sparse at best.
I still think it would be useful to stabilize that interface further, which would potentially open up a whole bunch of likely useful analysis & optimization. Maybe things are better in the future, there were some interesting talks in that direction (not refinement types in particular) at this years’ juliacon (specifically, this and this talk have given me some ideas for projects).
All that being said - I don’t think adopting refinement types at the core level & by default would be a boon. It’s just way too costly for the sideeffect-happy, non-trivial, non-functional code that is idiomatic in Julia. IMO it’d be more useful as a distinct check as part of a testsuite.